This paper proposes a novel decoding strategy, Active Layer-Contrastive Decoding (ActLCD), to address the hallucination problem in large-scale language models (LLMs). Unlike existing token-based methods, ActLCD uses a reinforcement learning-based policy to dynamically determine when to apply contrastive learning layers during the generation process. This approach views the decoding process as a sequential decision-making problem and optimizes fact accuracy through a reward-aware classifier. Experimental results show that ActLCD outperforms state-of-the-art methods across five benchmarks.
